iris-hostgl: flush presents don't wait, and come at most once a frame

[atomchild411]

iris-hostgl: flush presents don't wait, and come at most once a frame

SGI's IRIS GL flushes after every primitive it draws into the front buffer. Under host GL each flush was a full present the program waited for, so flight's cockpit drew itself instrument by instrument, over a second or more. On the board, and under SGI's own libraries on the emulated Max IMPACT, it appears at once.

The change

1. Swap intervals pace swaps only. GLX_SGI_swap_control applied its wait to every OP_SWAP: showing the front buffer, or a colour-index window, after a flush slept up to interval frames each. flight asks for swapinterval(2), so it drew a few primitives a second. A new bit 4 in the seventh slot marks a present as a flush, for colour-index windows, which have no front bit. Only swaps wait now.
2. Flush presents never wait for the presenter. Each one replaces that window's frame still waiting, so the screen gets the newest drawing as fast as the presenter can put it there, and always the last one. Colour-index frames take the queue like RGBA ones. The queue keeps one waiting frame per window rather than one overall, so one window's flushes never drop another window's frame.
3. OP_PRESENT_LATER (0x1_0012). It takes OP_SWAP's slots.
   • On the GL thread it is answered 1 at once and queued in order with the client's buffers. The GL thread shows a client's window at most once per FLUSH_GAP (a 60 Hz frame).
   • When a present is still waiting, it is shown when the gap is up; the thread wakes for that when nothing else comes. It is also shown before the client's next call that isn't a buffer, which is the only thing that can change what the client draws to.
   • Without the GL thread it is OP_SWAP; a host without it answers EINVAL.
   • The library sends it only for a window whose last present the display took, so nothing ever has to come back.
   • OP_SWAP's body moves into GlService::swap, unchanged, for both.

Tests

• Unit tests: flush_presents_wait_out_the_gap, present_later_is_answered_at_once_and_runs_in_order. The first sets its own gap (GlService::flush_gap), so it doesn't depend on how fast the host draws. With FLUSH_GAP itself it failed about one run in six on a loaded VM, where a present took longer than a frame. It passed 10 of 10 there afterwards.
• cargo test --manifest-path iris-hostgl/Cargo.toml: 35 passed on macOS.
• cargo test --lib mgras: 177 passed.
• flight under host GL (IRIX 6.5.22m, Max IMPACT, macOS host, guest tools v20261017-proto2), measured from screenshots every 0.1 s from launch:
  • cockpit complete in 0.56–0.66 s on this branch;
  • 2 s on upstream IRIS with the same guest library;
  • 0.55 s under SGI's own libraries on the emulated board.
• gltest all ok (n32, o32), glcheck PASSED, irisgltest the same 12 failures as SGI's libraries.
• On Linux (Debian 12 arm64 VM, Mesa, virgl):
  • cargo test: 35 passed.
  • flight's cockpit complete in 2.24 s. Upstream IRIS on the same VM had drawn only a few instruments after 6 s. SGI's own libraries take 1.45 s there; the VM's GPU is the difference.
  • Measured together with iris-hostgl: SGI_color_matrix's state on a host without a colour matrix (Linux) #246, which SGI's IRIS GL needs on Linux.

🤖 Generated with Claude Code